By the end of 2028, OpenAI and Anthropic could be running the majority of compute worldwide that is actually put to work, SemiAnalysis founder Dylan Patel told podcast host Dwarkesh Patel. Patel’s evidence starts with this year’s buildout: both labs started 2026 at roughly 2 gigawatts each and are on track to pass 5 gigawatts each by year end, a share he pegs at about 30 percent of all newly added compute. The stakes go beyond hardware allocation: if the trend holds, a handful of companies would set the price and pace of AI progress for everyone else.
Patel’s argument rests on a revenue mechanism, not a market-share prediction. Anthropic now pulls in up to $50 million of revenue for every megawatt of compute it runs, he said, against a base cost of roughly $10 to $15 million per megawatt, a margin wide enough to let the labs outbid other buyers for scarce chips and power. That gap, he argued, is what pulls compute toward OpenAI and Anthropic rather than toward hyperscalers or smaller labs: anyone bidding for the same gigawatt has to match a return the labs can already generate.
Next year’s contracts, already signed, push that figure to 40 to 50 percent of all new compute added, Patel said, up from about 30 percent now. His timeline then compresses further: OpenAI and Anthropic together absorb half of every new watt added globally by the close of 2027, and by the end of 2028 that climbs to a 70 to 80 percent grip on incremental supply. Layer in the efficiency gains from newer chip generations, and Patel’s math points to the two labs holding the working majority of the planet’s compute capacity outright.
Patel’s own numbers complicate the forecast he offers. He said global AI capital expenditure will pass $2 trillion by 2028, while labs generating hundreds of billions in revenue next year still cannot fund that buildout from cash flow alone, which means outside capital and hyperscaler partners remain part of the picture regardless of who ends up running the workloads. He also pointed to a countervailing force: regulatory and safety reviews that have already delayed releases, such as OpenAI holding back a model reportedly code-named Astra and Anthropic sitting on an unreleased system it internally calls “Model 2,” which Patel says is the next Mythos generation. If that pattern continues, he said, the labs’ revenue-per-megawatt advantage stalls, since competitors catch up while the frontier labs sit on unreleased capability.
The conversation also covered a downside case Patel did not fully resolve: whether the buildout scales into a sovereign debt problem, with hyperscaler borrowing pushing up interest rates broadly enough to strain economies with little AI exposure and pressure non-AI equities. The working assumption behind that risk, Patel said, is total AI capital spending topping $10 trillion before this decade closes, though he stopped short of putting a probability on the crisis scenario itself.
Patel’s framing does not include independent verification of the specific gigawatt or margin figures, which come from his own compute-market research at SemiAnalysis rather than from public filings by OpenAI or Anthropic. For operators building on either company’s models, the practical read is that pricing power, not just model quality, is now a lab moat: a customer’s negotiating leverage on inference costs may shrink as the compute gap widens, making multi-vendor contracts and open-weight fallback options worth pricing in before 2027 renewal cycles.
Reported by The Dwarkesh Podcast, episode with Dylan Patel published 25 August 2026.
